Skip to content

local model

LocalModel Module

This module provides the LocalModel class that allows loading, inference, and benchmark testing of models in a local environment. It supports detection and segmentation tasks, and utilizes ONNXRuntime for model execution.

Classes:

Name Description
LocalModel

A class for managing and interacting with local models.

Functions:

Name Description
__init__

Initializes the LocalModel instance, loading the model, metadata, and setting up the runtime.
_read_metadata

Reads the model metadata from a JSON file.
_annotate

Annotates the input image with detection or segmentation results.
infer

Runs inference on an input image, with optional annotation.
benchmark

Benchmarks the model's inference performance over a specified number of iterations and input size.

LocalModel #

Source code in focoos/local_model.py 
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
class LocalModel:
    def __init__(
        self,
        model_dir: Union[str, Path],
        runtime_type: Optional[RuntimeTypes] = None,
    ):
        """
        Initialize a LocalModel instance.

        This class sets up a local model for inference by initializing the runtime environment,
        loading metadata, and preparing annotation utilities.

        Args:
            model_dir (Union[str, Path]): The path to the directory containing the model files.
            runtime_type (Optional[RuntimeTypes]): Specifies the runtime type to use for inference.
                Defaults to the value of `FOCOOS_CONFIG.runtime_type` if not provided.

        Raises:
            ValueError: If no runtime type is provided and `FOCOOS_CONFIG.runtime_type` is not set.
            FileNotFoundError: If the specified model directory does not exist.

        Attributes:
            model_dir (Union[str, Path]): Path to the model directory.
            metadata (ModelMetadata): Metadata information for the model.
            model_ref: Reference identifier for the model obtained from metadata.
            label_annotator (sv.LabelAnnotator): Utility for adding labels to the output,
                initialized with text padding and border radius.
            box_annotator (sv.BoxAnnotator): Utility for annotating bounding boxes.
            mask_annotator (sv.MaskAnnotator): Utility for annotating masks.
            runtime (ONNXRuntime): Inference runtime initialized with the specified runtime type,
                model path, metadata, and warmup iterations.

        The method verifies the existence of the model directory, reads the model metadata,
        and initializes the runtime for inference using the provided runtime type. Annotation
        utilities are also prepared for visualizing model outputs.
        """
        runtime_type = runtime_type or FOCOOS_CONFIG.runtime_type

        logger.debug(f"Runtime type: {runtime_type}, Loading model from {model_dir},")
        if not os.path.exists(model_dir):
            raise FileNotFoundError(f"Model directory not found: {model_dir}")
        self.model_dir: Union[str, Path] = model_dir
        self.metadata: ModelMetadata = self._read_metadata()
        self.model_ref = self.metadata.ref
        self.label_annotator = sv.LabelAnnotator(text_padding=10, border_radius=10)
        self.box_annotator = sv.BoxAnnotator()
        self.mask_annotator = sv.MaskAnnotator()
        self.runtime: ONNXRuntime = get_runtime(
            runtime_type,
            str(os.path.join(model_dir, "model.onnx")),
            self.metadata,
            FOCOOS_CONFIG.warmup_iter,
        )

    def _read_metadata(self) -> ModelMetadata:
        """
        Reads the model metadata from a JSON file.

        Returns:
            ModelMetadata: Metadata for the model.

        Raises:
            FileNotFoundError: If the metadata file does not exist in the model directory.
        """
        metadata_path = os.path.join(self.model_dir, "focoos_metadata.json")
        return ModelMetadata.from_json(metadata_path)

    def _annotate(self, im: np.ndarray, detections: sv.Detections) -> np.ndarray:
        """
        Annotates the input image with detection or segmentation results.

        Args:
            im (np.ndarray): The input image to annotate.
            detections (sv.Detections): Detected objects or segmented regions.

        Returns:
            np.ndarray: The annotated image with bounding boxes or masks.
        """
        classes = self.metadata.classes
        labels = [
            f"{classes[int(class_id)] if classes is not None else str(class_id)}: {confid * 100:.0f}%"
            for class_id, confid in zip(detections.class_id, detections.confidence)  # type: ignore
        ]
        if self.metadata.task == FocoosTask.DETECTION:
            annotated_im = self.box_annotator.annotate(scene=im.copy(), detections=detections)

            annotated_im = self.label_annotator.annotate(scene=annotated_im, detections=detections, labels=labels)
        elif self.metadata.task in [
            FocoosTask.SEMSEG,
            FocoosTask.INSTANCE_SEGMENTATION,
        ]:
            annotated_im = self.mask_annotator.annotate(scene=im.copy(), detections=detections)
        return annotated_im

    def infer(
        self,
        image: Union[bytes, str, Path, np.ndarray, Image.Image],
        threshold: float = 0.5,
        annotate: bool = False,
    ) -> Tuple[FocoosDetections, Optional[np.ndarray]]:
        """
        Run inference on an input image and optionally annotate the results.

        Args:
            image (Union[bytes, str, Path, np.ndarray, Image.Image]): The input image to infer on.
                This can be a byte array, file path, or a PIL Image object, or a NumPy array representing the image.
            threshold (float, optional): The confidence threshold for detections. Defaults to 0.5.
                Detections with confidence scores below this threshold will be discarded.
            annotate (bool, optional): Whether to annotate the image with detection results. Defaults to False.
                If set to True, the method will return the image with bounding boxes or segmentation masks.

        Returns:
            Tuple[FocoosDetections, Optional[np.ndarray]]: A tuple containing:
                - `FocoosDetections`: The detections from the inference, represented as a custom object (`FocoosDetections`).
                This includes the details of the detected objects such as class, confidence score, and bounding box (if applicable).
                - `Optional[np.ndarray]`: The annotated image, if `annotate=True`.
                This will be a NumPy array representation of the image with drawn bounding boxes or segmentation masks.
                If `annotate=False`, this value will be `None`.

        Raises:
            ValueError: If the model is not deployed locally (i.e., `self.runtime` is `None`).
        """
        assert self.runtime is not None, "Model is not deployed (locally)"
        resize = None  #!TODO  check for segmentation
        if self.metadata.task == FocoosTask.DETECTION:
            resize = 640 if not self.metadata.im_size else self.metadata.im_size
        logger.debug(f"Resize: {resize}")
        t0 = perf_counter()
        im1, im0 = image_preprocess(image, resize=resize)
        t1 = perf_counter()
        detections = self.runtime(im1.astype(np.float32), threshold)
        t2 = perf_counter()
        if resize:
            detections = scale_detections(detections, (resize, resize), (im0.shape[1], im0.shape[0]))
        logger.debug(f"Inference time: {t2 - t1:.3f} seconds")
        im = None
        if annotate:
            im = self._annotate(im0, detections)

        out = sv_to_focoos_detections(detections, classes=self.metadata.classes)
        t3 = perf_counter()
        out.latency = {
            "inference": round(t2 - t1, 3),
            "preprocess": round(t1 - t0, 3),
            "postprocess": round(t3 - t2, 3),
        }
        return out, im

    def benchmark(self, iterations: int, size: int) -> LatencyMetrics:
        """
        Benchmark the model's inference performance over multiple iterations.

        Args:
            iterations (int): Number of iterations to run for benchmarking.
            size (int): The input size for each benchmark iteration.

        Returns:
            LatencyMetrics: Latency metrics including time taken for inference.
        """
        return self.runtime.benchmark(iterations, size)

__init__(model_dir, runtime_type=None) #

Initialize a LocalModel instance.

This class sets up a local model for inference by initializing the runtime environment, loading metadata, and preparing annotation utilities.

Parameters:

Name Type Description Default
model_dir Union[str, Path]

The path to the directory containing the model files.

 required
runtime_type Optional[RuntimeTypes]

Specifies the runtime type to use for inference. Defaults to the value of FOCOOS_CONFIG.runtime_type if not provided.

 None

Raises:

Type Description
ValueError

If no runtime type is provided and FOCOOS_CONFIG.runtime_type is not set.
FileNotFoundError

If the specified model directory does not exist.

Attributes:

Name Type Description
model_dir Union[str, Path]

Path to the model directory.
metadata ModelMetadata

Metadata information for the model.
model_ref ModelMetadata

Reference identifier for the model obtained from metadata.
label_annotator LabelAnnotator

Utility for adding labels to the output, initialized with text padding and border radius.
box_annotator BoxAnnotator

Utility for annotating bounding boxes.
mask_annotator MaskAnnotator

Utility for annotating masks.
runtime ONNXRuntime

Inference runtime initialized with the specified runtime type, model path, metadata, and warmup iterations.

The method verifies the existence of the model directory, reads the model metadata, and initializes the runtime for inference using the provided runtime type. Annotation utilities are also prepared for visualizing model outputs.

Source code in focoos/local_model.py 
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
def __init__(
    self,
    model_dir: Union[str, Path],
    runtime_type: Optional[RuntimeTypes] = None,
):
    """
    Initialize a LocalModel instance.

    This class sets up a local model for inference by initializing the runtime environment,
    loading metadata, and preparing annotation utilities.

    Args:
        model_dir (Union[str, Path]): The path to the directory containing the model files.
        runtime_type (Optional[RuntimeTypes]): Specifies the runtime type to use for inference.
            Defaults to the value of `FOCOOS_CONFIG.runtime_type` if not provided.

    Raises:
        ValueError: If no runtime type is provided and `FOCOOS_CONFIG.runtime_type` is not set.
        FileNotFoundError: If the specified model directory does not exist.

    Attributes:
        model_dir (Union[str, Path]): Path to the model directory.
        metadata (ModelMetadata): Metadata information for the model.
        model_ref: Reference identifier for the model obtained from metadata.
        label_annotator (sv.LabelAnnotator): Utility for adding labels to the output,
            initialized with text padding and border radius.
        box_annotator (sv.BoxAnnotator): Utility for annotating bounding boxes.
        mask_annotator (sv.MaskAnnotator): Utility for annotating masks.
        runtime (ONNXRuntime): Inference runtime initialized with the specified runtime type,
            model path, metadata, and warmup iterations.

    The method verifies the existence of the model directory, reads the model metadata,
    and initializes the runtime for inference using the provided runtime type. Annotation
    utilities are also prepared for visualizing model outputs.
    """
    runtime_type = runtime_type or FOCOOS_CONFIG.runtime_type

    logger.debug(f"Runtime type: {runtime_type}, Loading model from {model_dir},")
    if not os.path.exists(model_dir):
        raise FileNotFoundError(f"Model directory not found: {model_dir}")
    self.model_dir: Union[str, Path] = model_dir
    self.metadata: ModelMetadata = self._read_metadata()
    self.model_ref = self.metadata.ref
    self.label_annotator = sv.LabelAnnotator(text_padding=10, border_radius=10)
    self.box_annotator = sv.BoxAnnotator()
    self.mask_annotator = sv.MaskAnnotator()
    self.runtime: ONNXRuntime = get_runtime(
        runtime_type,
        str(os.path.join(model_dir, "model.onnx")),
        self.metadata,
        FOCOOS_CONFIG.warmup_iter,
    )

benchmark(iterations, size) #

Benchmark the model's inference performance over multiple iterations.

Parameters:

Name Type Description Default
iterations int

Number of iterations to run for benchmarking.

 required
size int

The input size for each benchmark iteration.

 required

Returns:

Name Type Description
LatencyMetrics LatencyMetrics

Latency metrics including time taken for inference.
Source code in focoos/local_model.py 
197
198
199
200
201
202
203
204
205
206
207
208
def benchmark(self, iterations: int, size: int) -> LatencyMetrics:
    """
    Benchmark the model's inference performance over multiple iterations.

    Args:
        iterations (int): Number of iterations to run for benchmarking.
        size (int): The input size for each benchmark iteration.

    Returns:
        LatencyMetrics: Latency metrics including time taken for inference.
    """
    return self.runtime.benchmark(iterations, size)

infer(image, threshold=0.5, annotate=False) #

Run inference on an input image and optionally annotate the results.

Parameters:

Name Type Description Default
image Union[bytes, str, Path, ndarray, Image]

The input image to infer on. This can be a byte array, file path, or a PIL Image object, or a NumPy array representing the image.

 required
threshold float

The confidence threshold for detections. Defaults to 0.5. Detections with confidence scores below this threshold will be discarded.

 0.5
annotate bool

Whether to annotate the image with detection results. Defaults to False. If set to True, the method will return the image with bounding boxes or segmentation masks.

 False

Returns:

Type Description
Tuple[FocoosDetections, Optional[ndarray]]

Tuple[FocoosDetections, Optional[np.ndarray]]: A tuple containing: - FocoosDetections: The detections from the inference, represented as a custom object (FocoosDetections). This includes the details of the detected objects such as class, confidence score, and bounding box (if applicable). - Optional[np.ndarray]: The annotated image, if annotate=True. This will be a NumPy array representation of the image with drawn bounding boxes or segmentation masks. If annotate=False, this value will be None.

Raises:

Type Description
ValueError

If the model is not deployed locally (i.e., self.runtime is None).
Source code in focoos/local_model.py 
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
def infer(
    self,
    image: Union[bytes, str, Path, np.ndarray, Image.Image],
    threshold: float = 0.5,
    annotate: bool = False,
) -> Tuple[FocoosDetections, Optional[np.ndarray]]:
    """
    Run inference on an input image and optionally annotate the results.

    Args:
        image (Union[bytes, str, Path, np.ndarray, Image.Image]): The input image to infer on.
            This can be a byte array, file path, or a PIL Image object, or a NumPy array representing the image.
        threshold (float, optional): The confidence threshold for detections. Defaults to 0.5.
            Detections with confidence scores below this threshold will be discarded.
        annotate (bool, optional): Whether to annotate the image with detection results. Defaults to False.
            If set to True, the method will return the image with bounding boxes or segmentation masks.

    Returns:
        Tuple[FocoosDetections, Optional[np.ndarray]]: A tuple containing:
            - `FocoosDetections`: The detections from the inference, represented as a custom object (`FocoosDetections`).
            This includes the details of the detected objects such as class, confidence score, and bounding box (if applicable).
            - `Optional[np.ndarray]`: The annotated image, if `annotate=True`.
            This will be a NumPy array representation of the image with drawn bounding boxes or segmentation masks.
            If `annotate=False`, this value will be `None`.

    Raises:
        ValueError: If the model is not deployed locally (i.e., `self.runtime` is `None`).
    """
    assert self.runtime is not None, "Model is not deployed (locally)"
    resize = None  #!TODO  check for segmentation
    if self.metadata.task == FocoosTask.DETECTION:
        resize = 640 if not self.metadata.im_size else self.metadata.im_size
    logger.debug(f"Resize: {resize}")
    t0 = perf_counter()
    im1, im0 = image_preprocess(image, resize=resize)
    t1 = perf_counter()
    detections = self.runtime(im1.astype(np.float32), threshold)
    t2 = perf_counter()
    if resize:
        detections = scale_detections(detections, (resize, resize), (im0.shape[1], im0.shape[0]))
    logger.debug(f"Inference time: {t2 - t1:.3f} seconds")
    im = None
    if annotate:
        im = self._annotate(im0, detections)

    out = sv_to_focoos_detections(detections, classes=self.metadata.classes)
    t3 = perf_counter()
    out.latency = {
        "inference": round(t2 - t1, 3),
        "preprocess": round(t1 - t0, 3),
        "postprocess": round(t3 - t2, 3),
    }
    return out, im